[VickyXAI]

Problem

On 2026-11-12, blockrun_chat on a Base wallet holding about $0.001 booked $0.104 across 3-4 zai/glm-5.3 calls (sessionSpend, and cost_usd with cost_is_estimate: false), but no payment settled. CDP verify failed, the gateway served the calls on its enterprise local-trust path, and its settlement after the stream then failed ([ALERT][enterprise-unpaid] for each call). One call ended in an upstream error and still came back as a success, with served_model: "error" and a final cost_usd: 0.04443.

Root cause

• src/tools/chat.ts:331 (withSettledCost, wallet branch) set costIsEstimate: d === null. In other words, any delta in the SDK's getSpending() counter was treated as settled. The SDK (@blockrun/llm 3.26.5, LLMClient.chatCompletionStream) adds the signed amount whenever a paid request comes back 2xx, unless the gateway names a skipped settlement. It never learns whether the settlement landed. On a stream it cannot know: the gateway sends the headers (just SSE_HEADERS, no receipt) before it settles. Line 335 did the same thing for failures and booked any delta as certainty: "settled" ("the charge stands").
• src/utils/chat-stream.ts ChatAccumulator.fold recognised only an { error } object. The gateway's sseError frame (model: "error", id: "error-<ts>", content "\n\n[Error: …]") was read as normal content, so a failed call came back as the answer.
• src/tools/chat-anthropic.ts:503 (the native thinking path) had the same problem: on the wallet rail it always set costIsEstimate = false.

The settled-vs-maybe split from 0.55.2/0.55.3 decided on how the call ended (it threw, or the edge returned 5xx). It never checked for a receipt, so a 2xx that was served but never settled always counted as settled.

Change

• New settlementTxFromHeaders(). It counts as proof only a receipt with a transaction: PAYMENT-RESPONSE/X-PAYMENT-RESPONSE with a non-empty transaction and no success: false, or X-Payment-Receipt. X-Payment-Settled: false overrides both. This matches the gateway's non-stream route, which writes success: true with an empty transaction plus X-Payment-Settled: false when its settlement fails. ChatOutcome and AcceptedThenFailedError now carry settlementTx. The non-streaming path reads the SDK's settlement.transaction. The Solana frame stream has no headers, so it is always unproven.
• Wallet rail: a cost is final (cost_is_estimate: false, settlement_tx set) only with that proof. Without it, cost_is_estimate: true and the reply says the amount was signed but not confirmed as settled. The ledger still books the signed amount, so global budgets and delegated limits stay conservative.
• A failure after a counter move with no receipt is booked as "unknown" (a maybe-charge). The routing loop still stops and the text still warns against retrying.
• The gateway's in-band error frame is now an AcceptedThenFailedError. Any partial text before it is kept.
• The native Claude path follows the same rule.

Paid streams never carry a receipt today, so paid wallet chat will normally report cost_is_estimate: true. That is what this server can actually observe. The gateway does not change.

Tests

• chat-wallet-after-payment.test.ts (Solana and Base):
  • (a) a successful paid stream with no receipt: still booked, cost_is_estimate: true, no settlement_tx.
  • (b) the gateway's model: "error" frame: isError, no served_model: "error", booked as a maybe, routing loop stops.
  • An error frame after some text keeps the partial text.
  • Base only: a receipt makes the cost final with settlement_tx; X-Payment-Settled: false plus an empty transaction is not proof; a failure that has a receipt still says "the charge stands".
• chat-anthropic-settled.test.ts: native success with and without a receipt, and a failure after connect with and without a receipt.
• Updated the existing "charge stands" assertions where there is no receipt, in chat-settled-on-error and the native partial-text test.
• I checked the new compat tests against the old src: 10 of them fail there.

npm run build, npm run typecheck and npm test all exit 0 (1491 tests, 1491 pass).

Not in this PR

blockrun_wallet sessionSpend is still one figure that includes maybe-charges, as the budget design intends. If a confirmed/unconfirmed split is wanted, it should cover every paid tool, not just chat. Other tools that book the SDK counter (image, music, speech, etc.) were not audited for the same rule.

🤖 Generated with Claude Code